lavc: Edge emulation with dst/src linesize

Allow supporting files for which the image stride is smaller than
the maximum block size + number of subpel mc taps, e.g. a 64x64 VP9
file or a 16x16 VP8 file with -fflags +emu_edge.

1 files changed, 160 insertions, 30 deletions

diff --git a/libavcodec/x86/videodsp_init.c b/libavcodec/x86/videodsp_init.c
index f7fd267e11..79d980194c 100644
--- a/libavcodec/x86/videodsp_init.c
+++ b/libavcodec/x86/videodsp_init.c
@@ -28,29 +28,127 @@
 #include "libavcodec/videodsp.h"
 
 #if HAVE_YASM
-typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src,
-                                x86_reg linesize, x86_reg start_y,
-                                x86_reg end_y, x86_reg block_h,
-                                x86_reg start_x, x86_reg end_x,
-                                x86_reg block_w);
-extern emu_edge_core_func ff_emu_edge_core_mmx;
-extern emu_edge_core_func ff_emu_edge_core_sse;
-
-static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
-                                              ptrdiff_t linesize,
-                                              int block_w, int block_h,
-                                              int src_x, int src_y,
-                                              int w, int h,
-                                              emu_edge_core_func *core_fn)
+typedef void emu_edge_vfix_func(uint8_t *dst, const uint8_t *src,
+                                x86_reg dst_stride, x86_reg src_stride,
+                                x86_reg start_y, x86_reg end_y, x86_reg bh);
+typedef void emu_edge_vvar_func(uint8_t *dst, const uint8_t *src,
+                                x86_reg dst_stride, x86_reg src_stride,
+                                x86_reg start_y, x86_reg end_y, x86_reg bh,
+                                x86_reg w);
+
+extern emu_edge_vfix_func ff_emu_edge_vfix1_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix2_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix3_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix4_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix5_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix6_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix7_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix8_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix9_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix10_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix11_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix12_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix13_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix14_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix15_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix16_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix17_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix18_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix19_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix20_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix21_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix22_mmx;
+#if ARCH_X86_32
+static emu_edge_vfix_func *vfixtbl_mmx[22] = {
+    &ff_emu_edge_vfix1_mmx,  &ff_emu_edge_vfix2_mmx,  &ff_emu_edge_vfix3_mmx,
+    &ff_emu_edge_vfix4_mmx,  &ff_emu_edge_vfix5_mmx,  &ff_emu_edge_vfix6_mmx,
+    &ff_emu_edge_vfix7_mmx,  &ff_emu_edge_vfix8_mmx,  &ff_emu_edge_vfix9_mmx,
+    &ff_emu_edge_vfix10_mmx, &ff_emu_edge_vfix11_mmx, &ff_emu_edge_vfix12_mmx,
+    &ff_emu_edge_vfix13_mmx, &ff_emu_edge_vfix14_mmx, &ff_emu_edge_vfix15_mmx,
+    &ff_emu_edge_vfix16_mmx, &ff_emu_edge_vfix17_mmx, &ff_emu_edge_vfix18_mmx,
+    &ff_emu_edge_vfix19_mmx, &ff_emu_edge_vfix20_mmx, &ff_emu_edge_vfix21_mmx,
+    &ff_emu_edge_vfix22_mmx
+};
+#endif
+extern emu_edge_vvar_func ff_emu_edge_vvar_mmx;
+extern emu_edge_vfix_func ff_emu_edge_vfix16_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix17_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix18_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix19_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix20_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix21_sse;
+extern emu_edge_vfix_func ff_emu_edge_vfix22_sse;
+static emu_edge_vfix_func *vfixtbl_sse[22] = {
+    ff_emu_edge_vfix1_mmx,  ff_emu_edge_vfix2_mmx,  ff_emu_edge_vfix3_mmx,
+    ff_emu_edge_vfix4_mmx,  ff_emu_edge_vfix5_mmx,  ff_emu_edge_vfix6_mmx,
+    ff_emu_edge_vfix7_mmx,  ff_emu_edge_vfix8_mmx,  ff_emu_edge_vfix9_mmx,
+    ff_emu_edge_vfix10_mmx, ff_emu_edge_vfix11_mmx, ff_emu_edge_vfix12_mmx,
+    ff_emu_edge_vfix13_mmx, ff_emu_edge_vfix14_mmx, ff_emu_edge_vfix15_mmx,
+    ff_emu_edge_vfix16_sse, ff_emu_edge_vfix17_sse, ff_emu_edge_vfix18_sse,
+    ff_emu_edge_vfix19_sse, ff_emu_edge_vfix20_sse, ff_emu_edge_vfix21_sse,
+    ff_emu_edge_vfix22_sse
+};
+extern emu_edge_vvar_func ff_emu_edge_vvar_sse;
+
+typedef void emu_edge_hfix_func(uint8_t *dst, x86_reg dst_stride,
+                                x86_reg start_x, x86_reg bh);
+typedef void emu_edge_hvar_func(uint8_t *dst, x86_reg dst_stride,
+                                x86_reg start_x, x86_reg n_words, x86_reg bh);
+
+extern emu_edge_hfix_func ff_emu_edge_hfix2_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix4_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix6_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix8_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix10_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix12_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix14_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix16_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix18_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix20_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix22_mmx;
+#if ARCH_X86_32
+static emu_edge_hfix_func *hfixtbl_mmx[11] = {
+    ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
+    ff_emu_edge_hfix8_mmx,  ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx,
+    ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_mmx, ff_emu_edge_hfix18_mmx,
+    ff_emu_edge_hfix20_mmx, ff_emu_edge_hfix22_mmx
+};
+#endif
+extern emu_edge_hvar_func ff_emu_edge_hvar_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix16_sse;
+extern emu_edge_hfix_func ff_emu_edge_hfix18_sse;
+extern emu_edge_hfix_func ff_emu_edge_hfix20_sse;
+extern emu_edge_hfix_func ff_emu_edge_hfix22_sse;
+static emu_edge_hfix_func *hfixtbl_sse[11] = {
+    ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
+    ff_emu_edge_hfix8_mmx,  ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx,
+    ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_sse, ff_emu_edge_hfix18_sse,
+    ff_emu_edge_hfix20_sse, ff_emu_edge_hfix22_sse
+};
+extern emu_edge_hvar_func ff_emu_edge_hvar_sse;
+
+static av_always_inline void emulated_edge_mc(uint8_t *dst, const uint8_t *src,
+                                              ptrdiff_t dst_stride,
+                                              ptrdiff_t src_stride,
+                                              x86_reg block_w, x86_reg block_h,
+                                              x86_reg src_x, x86_reg src_y,
+                                              x86_reg w, x86_reg h,
+                                              emu_edge_vfix_func **vfix_tbl,
+                                              emu_edge_vvar_func *v_extend_var,
+                                              emu_edge_hfix_func **hfix_tbl,
+                                              emu_edge_hvar_func *h_extend_var)
 {
-    int start_y, start_x, end_y, end_x, src_y_add = 0;
+    x86_reg start_y, start_x, end_y, end_x, src_y_add = 0, p;
+
+    if (!w || !h)
+         return;
 
     if (src_y >= h) {
-        src_y_add = h - 1 - src_y;
-        src_y     = h - 1;
+        src  -= src_y * src_stride;
+        src_y = src_y_add = h - 1;
     } else if (src_y <= -block_h) {
-        src_y_add = 1 - block_h - src_y;
-        src_y     = 1 - block_h;
+        src  -= src_y*src_stride;
+        src_y = src_y_add = 1 - block_h;
     }
     if (src_x >= w) {
         src   += w - 1 - src_x;
@@ -68,30 +166,62 @@ static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src,
     assert(start_y < end_y && block_h > 0);
 
     // fill in the to-be-copied part plus all above/below
-    src += (src_y_add + start_y) * linesize + start_x;
-    buf += start_x;
-    core_fn(buf, src, linesize, start_y, end_y,
-            block_h, start_x, end_x, block_w);
+    src += (src_y_add + start_y) * src_stride + start_x;
+    w = end_x - start_x;
+    if (w <= 22) {
+        vfix_tbl[w - 1](dst + start_x, src,
+                        dst_stride, src_stride,
+                        start_y, end_y, block_h);
+    } else {
+        v_extend_var(dst + start_x, src, dst_stride, src_stride,
+                     start_y, end_y, block_h, w);
+    }
+
+    // fill left
+    if (start_x) {
+        if (start_x <= 22) {
+            hfix_tbl[(start_x - 1) >> 1](dst, dst_stride, start_x, block_h);
+        } else {
+            h_extend_var(dst, dst_stride,
+                         start_x, (start_x + 1) >> 1, block_h);
+        }
+    }
+
+    // fill right
+    p = block_w - end_x;
+    if (p) {
+        if (p <= 22) {
+            hfix_tbl[(p - 1) >> 1](dst + end_x - (p & 1), dst_stride,
+                                   -!(p & 1), block_h);
+        } else {
+            h_extend_var(dst + end_x - (p & 1), dst_stride,
+                         -!(p & 1), (p + 1) >> 1, block_h);
+        }
+    }
 }
 
 #if ARCH_X86_32
 static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
-                                             ptrdiff_t linesize,
+                                             ptrdiff_t buf_stride,
+                                             ptrdiff_t src_stride,
                                              int block_w, int block_h,
                                              int src_x, int src_y, int w, int h)
 {
-    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
-                     w, h, &ff_emu_edge_core_mmx);
+    emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
+                     src_x, src_y, w, h, vfixtbl_mmx, &ff_emu_edge_vvar_mmx,
+                     hfixtbl_mmx, &ff_emu_edge_hvar_mmx);
 }
 #endif
 
-static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
-                                             ptrdiff_t linesize,
+static av_noinline void emulated_edge_mc_sse(uint8_t * buf,const uint8_t *src,
+                                             ptrdiff_t buf_stride,
+                                             ptrdiff_t src_stride,
                                              int block_w, int block_h,
                                              int src_x, int src_y, int w, int h)
 {
-    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
-                     w, h, &ff_emu_edge_core_sse);
+    emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
+                     src_x, src_y, w, h, vfixtbl_sse, &ff_emu_edge_vvar_sse,
+                     hfixtbl_sse, &ff_emu_edge_hvar_sse);
 }
 #endif /* HAVE_YASM */